Split Federated Learning-Enabled Deep Q-Networks for Generalized Path Planning in Distributed IoT Edge Platform

Internet of Things (IoT) devices and edge-platform robots require privacy-aware path planning under strict compute and bandwidth constraints, where continual on-device learning from real-world data is desirable but costly. We propose and evaluateSplit Federated Learning (SFL)–enabledDeep Q-Networks (DQN), termedSFL-DQN, which partitions the Q-network into a lightweight on-device encoder and a shared server-side action–value head with online/target stabilization. Devices keep raw observations local and exchange only intermediate feature representations and their gradients; periodicFederated Averaging (FedAvg)realigns client encoders to mitigate representation drift under heterogeneous, non-independent and identically distributed (non-IID) experience. In grid-world navigation withonemoving obstacle, SFL-DQN matches a non-federated, non-split DQN baseline in episodic return and task success and attainsBreadth-First Search (BFS)–equivalent shortest-path optimality on held-out maps. WithKclients, each device uses only 1/Kof the single-device interaction budget (e.g., 1,000 vs. 4,000 episodes whenK=4), reducing theclient-sidetraining burden by offloading the value-learning head to the server. Split–federated training introduces feature-level communication overhead and does not necessarily reduce system-wide energy under all network conditions; instead, SFL-DQN enables a practical edge–cloud learning loop without raw-data aggregation.